Session S34.6

Recognition of Cardiac Arrhythmias by Means of Beats Clustering on ECG-Holter Records

F Jiménez*, JL Rodríguez, E Delgado, D Cuesta, G Castellanos

Universidad Politécnica de Valencia
Valencia, Spain

A useful procedure for following the progress of cardiac diseases is the analysis on ECG-holter records. The automatic beats clustering of a ECG-holter record, is very useful due to the high number of beats (hundreds of miles), which implies high computational cost that in many cases saturates the dynamic memory of the processing equipment. On the other hand, the adjustment of the number of clusters for a certain record requires high precision, and the noise promotes the misalignment of this variable. The main goal of this work is to develop an algorithm for clustering beats in order to reduce the analysis time of a ECG-holter record. Thus, the analysis time is reduced to the revision of groups of morphologically similar beats, oriented to the recognition of cardiac arrhythmias. The proposed methodology consists in carrying out signal pre-processing by filtration of the baseline noise and high frequency noise. The location of the R peak is estimated by an algorithm based on nonlinear transformations and adaptive thresholding in space and time. The beat is extracted and a temporary normalization procedure is carried out using trace segmentation. For clustering beats, a representation based on Singular Value Decomposition (SVD) is made in order to adjust the existent number of beat clusters in the record, and to additionally take the reduced representation for each beat, since the computational cost increases exponentially as the number of points of each signal increases. A modification was made to the k-means algorithm with the purpose of optimizing the calculation of the centroids for each iteration. Finally, the set of beat clusters is obtained. The performance of the system is evaluated by the clustering error, knowing a priori the labels of the beats. The records used in this work belong to the MIT data base, which were acquired with a 360 Hz sampling frequency, 11 bits resolution and 10 mV range. The method has an efficiency of 85.6% and considering pathological versus normal beats, a sensitivity of 99.45% and specificity of 87.27%. This implies that few abnormal beats will be considered as normal, which is desirable. The fact of considering normal beats as possibly pathological means that the cardiologist will have to visually verify them, but this is not serious problem. On the other hand, the use of trace segmentation improves the efficiency index and reduces outliers.

(Abstract Control Number: 67)